This week, we dive into the main categories of machine learning algorithms and look at the pros and cons of each. For a brief definition of machine learning, check out last week's blog.
Three Categories of Machine Learning:
In machine learning, there are several algorithms you can use to model your data based on your use case. Most of them fall into three categories: supervised learning, unsupervised learning, and reinforcement learning. Each of them has different benefits in different situations, depending on the nature of the problem and the desired outcome.
Supervised Learning:
In supervised learning, both the desired input and the desired output are provided, and the machine must learn to map the former to the latter. To do this, the machine is trained using a statistically representative set of sample inputs and corresponding outputs.
An example of this might be teaching a machine to recognize the image of a dog. You would train the machine by showing it pictures of different breeds of dogs labeled as dogs versus pictures of cats labeled as cats. When it encounters the image of a dog, it recognizes it as a dog based on the data it was trained on. To do this, the specific features or characteristics of the input image are calculated and compared to the characteristics of the labeled images or objects.
An advantage of this approach is that the system is more controllable and the accuracy generally increases with the number of examples or labeled models provided. On the other hand, qualified persons must identify the examples or templates used for training. This can be labor intensive and time consuming and there are limits to the scalability of this approach.
Unsupervised Learning:
In unsupervised learning, the machine is not provided with labeled examples or previous patterns on which to base the analysis of input data. The machine must discover patterns and draw conclusions on its own without having the right answers. You classify or group data by discovering the similarity of features yourself. Through unsupervised learning, the machine would receive millions of images of dogs without labeling them as dogs. To decipher the clues, he used the text of the web copy or captions, noting in particular that the word "dog" appeared frequently in the various texts, and captioned the photos "dogs".
An advantage of this is that you don't need someone to label the examples or models and therefore no one is involved in the training. It can also be a scam because there is no human interaction to train the machine and it does not initially know whether the classifications it has made are correct or incorrect. Initially, the results may be erroneous. Patterns and groups discovered may or may not be useful to you. This too can be an advantage or a disadvantage. You may discover trends that you weren't looking for, but you may also not get the results you want.
Reinforcement Learning:
Reinforcement learning is distinguished from supervised and unsupervised learning by the continuous improvement of its model based on feedback. He learns by trial and error, the consequences of his actions and new decisions. As an action is performed, the success of the outcome is evaluated and assigned a positive or negative value. The algorithm strives for positive reviews and the model is trained on continuous feedback. A conceptual example of this might be a self-driving car that would receive a positive rating without getting from place to place without an accident.
The advantage of reinforcement learning is that there is a balance between trying what has worked in the past and trying new things to find new improvements. This means that the algorithm is likely to try new stocks or rankings in an incremental format and is more likely to discover new ideas and ways of doing things. Standard supervised learning algorithms cannot achieve this balance. A potential downside could be that you can't embed explicit rules later, as is the case with supervised learning (e.g. stopping at a red light), and a large amount of data can be necessary for the machine to provide the correct feedback. . Implementing reinforcement learning can also be quite difficult and requires a lot of specialized knowledge.
The right thing for you:
Each type of machine learning has advantages and disadvantages, and your success depends on the problem you want to solve and the results you want to achieve. The important thing is to understand your goals, then identify the types of data you need, and finally validate the analysis algorithms you choose.
Most Common Types of Machine Learning Problems
| Problem types | Details | Algorithms |
| Regression | When it comes to predicting numerical values, these problems are called regression problems. For example, predicting real estate prices | Linear regression, K-NN, random forest, neural networks |
| Classification | When data needs to be classified into different classes, it is called a classification problem. When there are two classes, it is called a binary classification problem. Many classes is a multinomial classification. For example, classifying whether a person has a disease or not. Classify whether a stock is "bought", "sold", or "held". Check out this related article: Machine Learning Techniques to Predict Stocks | Logistic regression, K-NN, gradient boosting classifier, neural networks |
| Clustering | When data points need to be classified into similar groups or clusters, it is called a clustering problem. | DBSCAN, Hierarchical clustering, Gaussian mixture models, BIRCH |
| Time-series forecasting | When there is a need to predict a number based on the time series data, it is called a time series prediction problem. A time series is a sequence of numeric data points in sequential order. Time series data means that the data is within a series of specific periods or time intervals. For example, a time series forecasting problem involves forecasting the sales demand for a product based on a set of input data, such as past sales, consumer sentiment, and weather. Another type of time series problem is demand forecasting. | LSTM, Exponential smoothing, Prophet, GARCH, TBATS, Dynamic linear models |
| Anomaly detection | When there are outliers to be found in the data set, the problem is called an anomaly detection problem. In other words, if a particular data set can be classified as an outlier or an unexpected event/element, it can be called an anomaly detection problem. For example, detecting fraudulent credit card transactions is an anomaly detection problem. | IsolationForest, Minimum covariance determinant, Local outlier factor, One-class SVM |
| Ranking | When it comes to ranking the results of a query or query based on certain criteria, the problem comes down to classification issues. We rank the query execution result based on the scores we assign to each result based on certain algorithms. These algorithms are called classification algorithms. Recommendation engines use the ranking algorithm to recommend the following articles. | (Bipartite Rankboost, Bipartite RankSVM) |
| Recommendation | If recommendations are needed, e.g. For example, "next item" to buy, "next video" to watch, or "next song" to listen to, the problem is known as recommendation problem. Solutions to such problems are called recommender systems. | Content-based and collaborative filtering machine learning methods |
| Data generation | When there is a need to generate data such as images, videos, articles, posts, etc., the problem is known as data generation problem. | Generative adversarial network (GAN), Hidden Markov models |
| Optimization | When it is necessary to generate a series of outputs that optimize the results with respect to a specific objective (objective function), the problem is called an objective function. | Linear programming methods, genetic programming |
Read More:
Christina Berglund